An MCP bridge that lets LLMs drive Photoshop directly

What it does

photoshop-mcp is a Model Context Protocol server that sits between an LLM assistant and a running Photoshop instance. You describe edits in plain language—remove backgrounds, resize for Instagram, batch-watermark a folder—and the AI translates the request into a sequence of Photoshop actions. The project bundles 122 discrete tools (106 atomic operations plus 16 pre-built recipes) into a single interface that works with Cursor, Claude, or a built-in browser chat window.

The interesting bit

The safety model is unusually thoughtful for glue code: every multi-step recipe is collapsed into a single undo step, and if Photoshop throws a structured error, the AI reads it and retries rather than blindly hammering the API. That means an LLM can experiment with portrait retouching or carousel slicing without turning your PSD into an unrecoverable mess.

Key highlights

• 122 tools covering masks, exports, generative fill, bulk watermarks, and social-media carousels.
• Recipes collapse complex workflows into one atomic, undoable Photoshop step.
• Works with Cursor, Claude Desktop, VS Code, or a standalone browser UI.
• Requires Photoshop (2012+) to be running locally on Windows or macOS.
• Unofficial and community-maintained; not affiliated with Adobe.

Caveats

• Anonymous usage analytics are collected by default and must be manually disabled.
• Adobe generative-fill features require a separate Adobe account.

Verdict

Worth a look if you batch-edit images in Photoshop and would rather delegate the clicking to an LLM. Skip it if you don’t already live in the Adobe ecosystem or if you need pixel-perfect, deterministic scripts rather than conversational trial-and-error.